Gemini 4 Argon
Runs MCP tools
Google’s first Gemini 4 model, released 13 October 2026. It is built for long, multi-step agent work, and Google reports the lowest prompt-injection success rate in its comparison set, which matters most when tools return text you did not write.
What Gemini 4 Argon is
Gemini 4 Argon is Google’s new frontier reasoning model, the successor to Gemini 3.1 Pro at the top of the line-up. It has a 1M-token window, up to 64K tokens of output on OpenRouter, and input across text, images, video, audio and PDFs. It thinks before it answers, with low, medium and high reasoning levels. Google trained it to keep context across long tasks, to pause for human confirmation before irreversible steps, and to report negative findings rather than paper over them. Artificial Analysis scores it 53 on its Intelligence Index at high reasoning, level with GPT-6 Astra and behind Claude Opus 5.5 at 58.
What Google DeepMind says
- AutomationBench (business workflows): 51.3%, against 42.5% for Claude Opus 5.5 and 41.4% for GPT-6 Astra.
- DeepSWE v1.1: 77.9%, against 74.2% for Claude Opus 5.5.
- Terminal-Bench 4.0: 57.4%, behind Claude Opus 5.5 at 66.4% and GPT-6 Astra at 58.2%.
- Gray Swan indirect prompt injection: 0.7% attack success across 15 adaptive attempts.
These are the vendor’s own claims, not measurements of ours. Run the model against your own server to find out whether they hold for your tools.
How it reaches MCP
- MCP Playground’s Agent Studio: paste your server URL above and run Gemini 4 Argon against it in the browser, with no Google key
- The Gemini API in Google AI Studio, whose Gen AI SDKs can pass an MCP session to the model as tools
- Google Antigravity, Google’s agent IDE, for Google AI Ultra subscribers
What to watch for
- Google’s strongest agent numbers are on business workflows (AutomationBench), and its weakest is Terminal-Bench, where it trails Claude Opus 5.5 by nine points. If your MCP server wraps a shell or a dev environment, do not read the workflow score as a promise.
- It is trained to stop and ask before irreversible steps. On a server with delete or payment tools that is what you want, but an unattended run can end on a confirmation question instead of a tool call. Count that as a pause, not a failure.
- OpenRouter lists tools, tool_choice, structured_outputs and seed, but not parallel_tool_calls. A client that expects several calls in one turn may see them one at a time.
Run the error-simulation mock and give Argon a task that hits simulate_partial_success, then the same task on Gemini 3.1 Pro. Google says Argon reports negative results honestly. Check whether it says which step failed or claims the whole task worked. All the mock servers →
Available here
| Model | Model ID |
|---|
| Gemini 4 Argon | google/gemini-4-argon |
Frequently asked questions
Does Gemini 4 Argon support tool calling?
Yes. OpenRouter lists tools, tool_choice and structured_outputs, and Google built it for autonomous multi-step agents. MCP servers work through any client that turns MCP tools into function calls, including the Gemini SDKs.
Gemini 4 Argon or Gemini 3.1 Pro?
Argon is the newer flagship and scores far higher on Artificial Analysis’s index (53 against about 30). Gemini 3.1 Pro has a published MCP Atlas score and Argon does not yet. Run the same task on both against your server and compare the calls.
Is Gemini 4 Argon safe to point at tools that return untrusted content?
Safer than most, by Google’s figure: 0.7% attack success on the Gray Swan indirect prompt-injection benchmark across 15 attempts. That is a vendor number, not zero, so keep permission checks on your side for tools that write or spend.
What reasoning levels does Gemini 4 Argon have?
Low, medium and high. On OpenRouter they map to reasoning_effort. Published benchmark scores were measured at high, so a run at low is not the model those numbers describe.